Async workers that keep the API responsive

HireGenix taught a simple rule: if resume parsing and interview evaluation share the request path with the API, the product feels broken the moment traffic arrives.

RabbitMQ workers with atomic job locking let background work fan out without double-processing. The API stays thin — accept the job, return status, move on.

Qdrant vector search for recommendations sat in the same spirit: keep expensive matching off the critical path, and expose results when they are ready.

The lesson I keep: concurrency is not a performance trick. It is how you protect the user-facing surface when AI work is slow and uneven.